Datasets:

SahmBenchmark
/

fatwa-training_standardized_new

---
license: apache-2.0
language:
- ar
tags:
- islamic-finance
- fatwa
- question-answering
- training
- instruction-tuning
- arabic
size_categories:
- 10K<n<100K
task_categories:
- question-answering
- text-generation
pretty_name: Fatwa Training Dataset (Standardized)
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
dataset_info:
  features:
  - name: id
    dtype: string
  - name: conversations
    list:
    - name: content
      dtype: string
    - name: role
      dtype: string
  - name: category
    dtype: string
  - name: is_referral
    dtype: string
  - name: question_length
    dtype: int64
  - name: answer_length
    dtype: int64
  splits:
  - name: train
    num_bytes: 15481402
    num_examples: 9953
  download_size: 6512899
  dataset_size: 15481402
---

# Fatwa Training Dataset (Standardized)

## Dataset Description

This dataset contains Islamic finance and jurisprudence fatwa question-answer pairs in a **standardized conversation format** for training Arabic language models. Each original sample has been augmented with **3 different prompt templates** to increase training diversity.

## Dataset Statistics

- **Total Samples**: 9,953
- **Unique Fatwas**: 6,212
- **Prompt Variations**: 3 per fatwa
- **Average Question Length**: 230.0 characters
- **Average Answer Length**: 493.6 characters

## Dataset Structure

### Data Fields

- `id`: Unique identifier for each fatwa (format: `fatwa_XXXXX`)
- `conversations`: List of conversation turns in chat format
  - `content`: The text content
  - `role`: Either "human" (question) or "agent" (answer)
- `category`: Islamic finance category
- `is_referral`: Whether the fatwa is mainly a referral (YES/NO)
- `question_length`: Character count of the original question
- `answer_length`: Character count of the answer

### Categories

- **zakat**: 4096 samples
- **riba**: 2047 samples
- **murabaha**: 1155 samples
- **gharar**: 711 samples
- **waqf**: 606 samples
- **ijara**: 469 samples
- **maysir**: 308 samples
- **musharaka**: 198 samples
- **mudharaba**: 188 samples
- **takaful**: 149 samples
- **sukuk**: 26 samples

### Prompt Templates

Each fatwa appears 3 times with different prompt styles:

1. **Formal Style**: "بناءً على أحكام الشريعة الإسلامية والفقه الإسلامي، أجب على السؤال التالي..."
2. **Concise Style**: "أجب على السؤال التالي وفقاً لأحكام الشريعة الإسلامية..."
3. **Expert Persona**: "أنت عالم متخصص في الفقه الإسلامي والمعاملات المالية..."

## Usage
```python
from datasets import load_dataset

dataset = load_dataset("SahmBenchmark/fatwa-training_standardized_new")

# Access training data
for example in dataset['train']:
    print(f"ID: {example['id']}")
    print(f"Human: {example['conversations'][0]['content']}")
    print(f"Agent: {example['conversations'][1]['content']}")
    print(f"Category: {example['category']}")
```

### For Fine-tuning
```python
from datasets import load_dataset

dataset = load_dataset("SahmBenchmark/fatwa-training_standardized_new")

def format_for_training(example):
    human_msg = example['conversations'][0]['content']
    agent_msg = example['conversations'][1]['content']
    return {"text": f"### Human: {human_msg}\n\n### Assistant: {agent_msg}"}

formatted_dataset = dataset.map(format_for_training)
```

## Categories

- **zakat**: Islamic almsgiving
- **riba**: Interest/usury-related rulings
- **murabaha**: Cost-plus financing
- **gharar**: Uncertainty in contracts
- **waqf**: Islamic endowment
- **ijara**: Islamic leasing
- **maysir**: Gambling-related rulings
- **musharaka**: Partnership financing
- **mudharaba**: Profit-sharing partnership
- **takaful**: Islamic insurance
- **sukuk**: Islamic bonds

## Citation
```bibtex
@dataset{fatwa_training_standardized,
  title={Fatwa Training Dataset (Standardized)},
  author={SahmBenchmark},
  year={2025},
  url={https://huggingface.co/datasets/SahmBenchmark/fatwa-training_standardized_new}
}
```

## License

Apache 2.0 License